Claude Code CLI

Claude Code is Anthropic's official AI coding agent — it runs in your terminal and lets you write, debug, and refactor code in natural language.

Step 1 — Install Node.js

Skip this if you already have it; check with node -v.

You need Node.js 18+ — download: https://nodejs.org/en/download

# Verify after installing
node -v
# A version number (e.g. v24.4.1) means it installed correctly

# During install, check "Automatically install the necessary tools"
# Reopen PowerShell after installing and verify
node -v

Step 2 — Install Claude Code

npm install -g @anthropic-ai/claude-code

npm install -g @anthropic-ai/claude-code

Verify the install:

claude --version

npm config set registry https://registry.npmmirror.com
npm install -g @anthropic-ai/claude-code

Step 3 — Configure the LMU AI API

Claude Code configures a custom API endpoint through settings.json.

Config file location

~/.claude/settings.json

C:\Users\YourUsername\.claude\settings.json

Create the config file

mkdir -p ~/.claude && touch ~/.claude/settings.json

mkdir "$env:USERPROFILE\.claude" -Force
New-Item "$env:USERPROFILE\.claude\settings.json" -Force

Write the config

Open settings.json in a text editor and add the following (replace the key with yours):

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.lmuai.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-your-lmu-ai-api-key",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
  }
}

Enabling the 1M context in Claude Code (optional)

Claude Opus 4.8 / Sonnet 5 support a 1M-token long context window (200K by default), useful for huge repos, long logs, and cross-file refactors. Add the [1M] suffix to the model ID to enable it.

Option A — manual /model switch (recommended for occasional use)

Keep a minimal settings.json (just ANTHROPIC_BASE_URL + ANTHROPIC_AUTH_TOKEN), start Claude Code, and type in the prompt:

/model claude-opus-5[1M]

or:

/model claude-sonnet-5[1M]

This switches to the 1M version with no config changes. It applies to the current session and reverts on exit.

Option B — a standing default in settings.json (recommended for daily long-context use)

Add two lines to the env section of settings.json so /model opus / /model sonnet default to the 1M version:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.lmuai.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-your-lmu-ai-api-key",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "claude-opus-5[1M]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "claude-sonnet-5[1M]",
    "API_TIMEOUT_MS": "3000000",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
  }
}

Save and restart Claude Code; selecting opus / sonnet in the /model command now uses the 1M version automatically.

Field reference

Using Chinese models (optional)

LMU AI supports Chinese LLMs (e.g. the Qwen series). By setting model in settings.json, you skip the manual /model switch on every launch and use the specified model directly.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.lmuai.com",
    "ANTHROPIC_AUTH_TOKEN": "sk-your-lmu-ai-api-key",
    "CLAUDE_CODE_ATTRIBUTION_HEADER": "0"
  },
  "model": "qwen3.8-max-preview",
  "effortLevel": "medium"
}

Field reference:

Step 4 — Launch Claude Code

Open a terminal in your project directory and run:

claude

No-approval mode (recommended)

claude --dangerously-skip-permissions

In this mode Claude Code runs commands automatically without confirming each step — handy inside a project directory.

Verify the config

After launching, type in the Claude Code interface:

/status

It shows the current:

• Model name
• API Base URL (should show https://api.lmuai.com)
• API key status

If the Base URL is correct, the config is working.

Common issues

401 Unauthorized

Cause: the API key is wrong, or the request still went to Anthropic's official endpoint.

Fix:

1. Confirm ANTHROPIC_AUTH_TOKEN in settings.json is your LMU AI key
2. Check whether a shell environment variable is overriding the config

echo $ANTHROPIC_BASE_URL

echo $env:ANTHROPIC_BASE_URL

3. Reopen the terminal, then launch Claude Code again

stream disconnected

Cause: an unstable local network, or a VPN / proxy / system proxy is on — a proxy rotating the IP breaks the connection.

Fix: turn off the VPN / proxy / system proxy and retry. LMU AI is a direct domestic connection and needs no VPN; a direct connection is fastest and most stable.

503 No available accounts

Cause: usually ~/.zshrc or ~/.bashrc sets global env vars like ANTHROPIC_AUTH_TOKEN / ANTHROPIC_BASE_URL, which override settings.json.

Fix: remove those lines from your shell config, or move them to a separate file you only source when launching Claude Code. See FAQ · Issue 6.

Response timeouts

Cause: the default timeout is short.

Fix: confirm API_TIMEOUT_MS is set to 3000000 (50 minutes) so long tasks are not cut off.

Tips

• Launch Claude Code from the project root so it picks up the project structure automatically
• Pass an instruction directly, e.g. claude "refactor this function for me"
• Ctrl+C interrupts the current task, /exit quits
• Type ? or /help to see all commands

Using /goal to make Claude Code work until the goal is met (optional)

/goal is a built-in Claude Code slash command (available since v2.1.139, May 2026) that sets a completion condition for the current session. Once set, Claude keeps going across turns on its own until the condition is judged met, rather than stopping when "it feels done." It is especially handy on LMU AI for long, run-to-completion tasks (large-repo refactors, implementing to an acceptance spec, clearing an issue backlog).

Older versions do not have this command. If it reports an unknown command, upgrade Claude Code to the latest version first.

Basic usage

A session has only one goal at a time; setting a new one replaces the old and starts a new turn immediately.

How it judges "met"

At the end of each turn, Claude Code hands your completion condition + this turn's conversation to a small, fast model (Haiku by default) to judge. The judge only looks at evidence already in the conversation — test output, build logs, file diffs — and will not re-run your whole CI behind your back. So write the condition in a form where Claude can "produce evidence" in the conversation.

The goal is cleared automatically in any of these cases:

• the condition is judged met;
• the model judges the condition impossible to satisfy;
• a turn hits an error that needs your intervention.

Writing a good goal condition

• Use a verifiable end state: e.g. "npm test exits 0", "tsc --noEmit reports no errors" — not subjective descriptions like "make the code prettier";
• Let evidence land in the conversation: print test / build results each turn so the judge model can see them;
• Scope it + cap the turns: e.g. "only change files under src/auth/, stop after at most 20 turns" to avoid spinning;
• Requires a trusted workspace (hooks enabled) to take effect.

Last updated: 2026-08-26